Liza

Liza coordinates AI coding agents for developers who want to delegate software work while keeping control over goals and constraints. It supports human-led pairing and multi-agent delivery runs that can proceed unattended.

• Goal-to-code delivery: Agents break down a goal, write specifications and code, review the results, and merge approved changes through enforced gates; the task graph can adapt during execution.
• Independent review: Separate author and reviewer roles assess artifacts throughout delivery, with software enforcing role boundaries, validation checks, and merge authority.
• Human steering: You can guide the work and add project guardrails while retaining authority over irreversible actions beyond the delegated workflow.
• Inspectable runs: Isolated Git worktrees separate concurrent changes, while a live terminal interface, shared task records, recorded prompts, and agent logs expose progress and decisions.

A shared behavioral contract requires agents to reason before acting, verify claims, disclose uncertainty, and respect scope. Crash recovery and context handoff support interrupted work, and a circuit breaker limits repeated failures.

Liza is open source under Apache 2.0. It supports Claude Opus through Claude Code and GPT-5 through Codex, using your existing provider subscriptions and coding CLI setup.